Data-center construction group
Division of a multinational electrical group · parent revenue in the tens of billions · Hundreds of bids a month · a ten-person estimating team
Drawing takeoff and ROM estimating for hyperscale data-center bids
From four issue-for-construction drawing sets, 190 sheets, agents itemized 1,045 units across 36 equipment families, matched the estimators' own counts on the big-ticket lines (88 chillers, 216 computer-room air handlers, 36 pumps), and surfaced a 161-versus-280 fan-wall-unit spread between drawings and proposal for adjudication. Hours, not the weeks a manual count takes.
- Organization
- Pre-construction estimating group inside a multinational electrical and building-systems company
- Bid set
- 4 drawing sets, 190 sheets, 3 project-manual volumes, the estimator's own markups
- Runs on
- Two dedicated GPUs: one for reasoning, one for visual takeoff
- Scope
- Takeoff, reconciliation, ROM estimate, proposal in the house format
Where they started.
Estimating is pre-construction's biggest bottleneck. The group quotes hundreds of bids a month with a ten-person estimating team, a single hyperscale data-center estimate can take one estimator two months, and what is visually on the printed page is what the subcontractor is on the hook for. The delta that hurts is schedules versus floor plans: a schedule says 420, the plans show 450, and the sub owns the difference. Every bid arrives as marked-up drawing sets and manual volumes with inconsistent or missing metadata, and nothing more is coming from the customer.
What was built.
- Agents read the drawings' own embedded data natively, page by page, and build a census in which a unit is its family, zone and number: a tag drawn on ten sheets is one unit, a range tag is twenty, and duplicate sightings across the clean and marked-up sets collapse to one.
- On this bid: 1,045 units across 36 equipment families, every count citing the sheets it was read from, cross-checked against the drawings' own schedule tables as a second, independent confirmation.
- The counts reconciled with the estimators' proposal on the lines that drive price: 88 air-cooled chillers, 216 computer-room air handlers across four GPU halls and the in-building hall, 36 secondary chilled-water pumps. Where drawings and proposal disagreed, 161 fan-wall units on the sheets against 280 in the proposal, the spread was surfaced for adjudication instead of averaged away.
- The rough-order-of-magnitude estimate renders in the team's own schedule-of-pricing format, with the same line items, alternates and totals their proposals already use.
- Estimators steer it in plain language: increase all labor by five percent, union state, twenty percent spares. Role-based guardrails cap how far an estimator can move labor or spares without an estimation manager, and finalizing an estimate waits for human approval.
- Specifications and manuals answer questions in the same conversation: what a division requires, which risks the estimator highlighted, what the RFI clarifications say.
What changed.
- A first-draft takeoff in hours against a manual count measured in weeks.
- Counts that reconcile with the team's own numbers, with sheet references anyone can open to check.
- Risk surfaced automatically: the 161-versus-280 spread is exactly the miss a subcontractor otherwise eats.
- Estimators spend their time on judgment, pricing and finesse instead of tallying.
- The same pipeline applies to the next bid set without re-engineering, and every correction trains the next run.
Bid drawings, the estimator's markups and the pricing behind a hyperscale data-center proposal are among the most competitively sensitive documents a company holds. Processing them on private inference keeps a bid inside the team that owns it.
Kirk-hosted to start, moving into the customer's Azure tenant without a rebuild
Components involved.
More like this.
Multi-brand industrial manufacturer
≈ $200M revenue across five brands (approx.) · 5 brands · 4 ERPs · 1 CRM
- Use Case:
- Consolidation, forecast and pricing intelligence across four ERPs
- Results:
- One governed lakehouse behind an AI agent: audit-grade consolidated financials that trace to the source transaction, forecast and pipeline by business unit, large-job margin watch, and plain-English inquiry over all of it, inside the company's own Azure tenant.
European logistics group
≈ $2B revenue · 6,000+ employees · 50+ countries
- Use Case:
- Quote control tower over forwarding systems
- Results:
- Quotes that track fuel prices, routing and political risk across the group's forwarding systems, assembled by agents over the company's own data lake and handed to the desk with the reasoning attached.
Flooring and concrete contractor
≈ $25M revenue (approx.) · 12 years of job history · Zoho CRM · Azure tenant
- Use Case:
- Estimation benchmarks from twelve years of costing sheets
- Results:
- Twelve years of costing sheets and work orders become a governed knowledge base; every new estimate is benchmarked by technology and size band against what past jobs actually cost, QA'd before approval, and driven from inside the CRM the sales team already lives in.
Have one like it?
Every spotlight started as a conversation about a process nobody liked doing, under a data-locality constraint.